Who wins the auction depends on what you measure
The idea came from a cloud prototype where you offer the capacity you don't use and buy what you need on top at the best price. We tested its market with agents that learn on their own.
It all started from a prototype of a cloud service. The idea is simple: every participant has their own machines, and most of the time part of their capacity sits idle. At another moment the same participant needs more than they own. Why shouldn’t the spare capacity go to someone who needs it right now, at a price the participants set themselves?
That way everyone gets the best price for the resources they take beyond their own, and a fair price for the ones they give. The heart of such a cloud is the market: how demand meets supply and who bids how.
Before building the cloud around the market, we wanted to know how the market itself would behave when the participants are represented by agents that learn on their own. In the paper “Multi-Agent Reinforcement Learning for Decentralized Computational Resource Negotiation in a Tokenized Double-Auction Market”, published in Mathematics in 2026, Georgi Tsochev, Adelina Aleksieva-Petrova and I test exactly that.
How we got to the question
One round of the market
Here is a simplified round with four buyers and four sellers. You are the buyer for whom the resource is worth 9. The others bid exactly what it is worth to them. All trades clear at one price: the midpoint between the last bid and the last ask that matched.
Start with a bid of 9, then lower it step by step. At first the price does not move. As soon as you are the last buyer who gets into a trade, a lower bid pulls the price down and you gain more. Lower it too far and you drop out of the trade, the resource goes to someone who values it less, and total welfare falls.
That is exactly what the agents do: they learn when to bid below their value. Good for one agent does not necessarily mean good for the market.
Who wins
The results below are the mean values over the 12 repetitions. Change the metric and watch the leader change too.
On reward and on social welfare MAPPO leads: 0.0140 [0.0124, 0.0154] for reward and 0.0952 [0.0854, 0.1045] for social welfare. On inequality, though, IQL does best, with the lowest Gini coefficient: 0.4477 [0.4360, 0.4613]. MAPPO, which leads on the other two metrics, has the highest coefficient of the four here: 0.579.
What follows from this
An algorithm that wins on one metric does not automatically win on all of them. Had we looked only at reward, we would have declared MAPPO the best and missed that with it the tokens concentrate in fewer hands.
That is why the contribution of the paper is a benchmark and an evaluation protocol: the same market, the same training budget, many repetitions, confidence intervals and several metrics reported together. The paper does not propose a new auction mechanism. It gives a way to compare any new algorithm fairly with the known ones.
The paper is open access: doi.org/10.3390/math14111828. All my publications are in Publications.